Fix load INC load weights compile error due to Transformer 4.45 upgrade.

[jiminha]

What does this PR do?

Error is caused by transformer PR integrated in 4.45 release: huggingface/transformers#33275
It removed unnecessary parameter called "loaded_state_dict_keys". However Synapse1.18 release is not compatible with this since it still uses "loaded_state_dict_keys", thus causing a compile error. For now we are overriding this method locally to fix the compile error temporarily. This will be fixed properly in Synapse 1.19 release.

Fixes # (issue)

SRAM_SLICER_SHARED_MME_INPUT_EXPANSION_ENABLED=false ENABLE_EXPERIMENTAL_FLAGS=1 \
python run_lm_eval.py \
-o acc_load_uint4_model.txt \
--model_name_or_path <model_path_in_hugging_face> \
--use_hpu_graphs \
--use_kv_cache \
--trim_logits \
--batch_size 1 \
--bf16 \
--attn_softmax_bf16 \
--bucket_size=128 \
--bucket_internal \
--load_quantized_model_with_inc

Above command will fail with compile error
TypeError: _load_state_dict_into_meta_model() got an unexpected keyword argument 'loaded_state_dict_keys'

[github-actions]

The code quality check failed, please run make style.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[jiminha]

Do not merge yet, I'm waiting for internal review of this code.

[MrGeva]

LGTM

[ftian1]

looks good to me

[xin3he]

LGTM

[xinhe3]

@jiminha I found that when transformer upgrade from 4.45.1 -> 4.45.2, another error will happen.
ValueError: HPUWeightOnlyLinear(in_features=11008, out_features=4096, bits=4, group_size=128, bias=True) does not have a parameter or a buffer named bias

we need another change like here.
"expected_keys": list(state_dict.keys()), -> expected_keys=self.loaded_state_dict_keys

[jiminha]

@jiminha I found that when transformer upgrade from 4.45.1 -> 4.45.2, another error will happen. ValueError: HPUWeightOnlyLinear(in_features=11008, out_features=4096, bits=4, group_size=128, bias=True) does not have a parameter or a buffer named bias

we need another change like here. "expected_keys": list(state_dict.keys()), -> expected_keys=self.loaded_state_dict_keys

This was not reproducible with run_generation command, but it seems this issue is seen with run_lm_eval. I will add this as well.
Is this change only required for this specific version, or is it backward compatible?